Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add strikersam/autonomous-ai-agency --skill training-stability-monitorgit clone --depth 1 https://github.com/strikersam/autonomous-ai-agencyWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/strikersam/autonomous-ai-agency/training-stability-monitor)<a href="https://agentmods.dev/skills/strikersam/autonomous-ai-agency/training-stability-monitor"><img src="https://agentmods.dev/badge/skills/strikersam/autonomous-ai-agency/training-stability-monitor/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/strikersam/autonomous-ai-agency/training-stability-monitor"><img src="https://agentmods.dev/badge/skills/strikersam/autonomous-ai-agency/training-stability-monitor.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00000 | $0.00753 |
| Opus 5 | $0.00000 | $0.00377 |
| Sonnet 5 | $0.00000 | $0.00151 |
| Haiku 4.5 | $0.00000 | $0.00075 |
Grade A, and why
training-stability-monitor scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 85 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Skill: training-stability-monitor
Purpose
Monitor and diagnose LLM/ML training instability — loss spikes, exploding gradients, dead neurons, and learning rate mismatches. Surfaces actionable fixes before they derail a training run.
Trigger
Use when:
- A training loss curve shows spikes or divergence
- Gradients are exploding or vanishing
- Model outputs collapse (repetition, gibberish, empty)
- You want a pre-flight check before a long training run
What It Does
- Scans training logs for loss spike signatures (sudden >2x jump in loss)
- Checks gradient norms — flags if norm exceeds configured threshold (default: 1.0)
- Validates LR schedule — warns if no warmup is configured, or if LR is too high for model size
- Checks batch size vs model size ratio — common source of instability
- Detects loss plateau — flat loss for N steps may indicate dead optimizer state
- Recommends recovery steps — rollback checkpoint, reduce LR, increase gradient clipping
Usage
/training-stability-monitor [log_file_or_directory] [--threshold 1.0] [--window 100]
Output Format
=== Training Stability Report ===
[PASS/WARN/FAIL] Loss Spike Detection
[PASS/WARN/FAIL] Gradient Norm
[PASS/WARN/FAIL] LR Schedule
[PASS/WARN/FAIL] Batch Size Ratio
[PASS/WARN/FAIL] Loss Plateau
CRITICAL ISSUES: N
WARNINGS: N
Recommended Actions:
1. ...
Key Lessons (from LLM-from-scratch practitioners)
- Loss spikes are normal but recoverable — the fix is gradient clipping + LR warmup, not stopping the run
- Gradient norm logging is non-negotiable — without it you're flying blind
- Warmup is mandatory — jumping to peak LR on step 0 destabilizes attention weights
- Save checkpoints every N steps — not just at epoch boundaries; spikes need rollback points
- BF16 > FP16 for stability on modern hardware; overflow is silent in FP16
Integration Points
- Works alongside
debug-tracerfor step-level tracing - Works alongside
insightsfor surfacing patterns across runs - Can feed into
session-handoffto carry stability context across sessions
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 85 lines · 0 tokens per session scan A 67cf24f26750
training-stability-monitor is a skill published in the GitHub repository strikersam/autonomous-ai-agency (8 stars, last pushed today), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 753 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
digital-health-clinical-asr-eval
Stage 3 of Clinical ASR Flywheel. Score a NeMo manifest, produce the five-section KER leaderboard (by-ipasource diagnostic). Not for ASR auth (/riva-asr).
digital-health-clinical-asr-finetune
Stage 4 of the Clinical ASR Flywheel. Use when priority KER is above 0.3 to run stock NeMo SFT on Parakeet TDT v2 and offline cycle N+1 re-eval. NOT for generic word boosting (use /finetune-asr).
nemo-mbridge-perf-expert-parallel-overlap
Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlapmoeexpertparallelcomm, delaywgradcompute, and flex dispatcher backends such as DeepEP and HybridEP.
i4h-workflow-dataset-annotate
Grade or filter workflow HDF5 episodes with an OpenAI-compatible vision model. Use for visual success labels; do not use for replay, policy evaluation, or recordings without frames.
i4h-workflow-dataset-convert
Convert workflow HDF5 recordings to LeRobot datasets for training or browser inspection. Use for conversion; do not use for replay, augmentation, or raw-data repair.
i4h-lerobot-viz
Serve and visually inspect a converted LeRobot dataset in the browser. Use for videos and state/action timelines; do not use for raw workflow HDF5 or incomplete conversion output.